Artificial Intelligence and the future of the accounting ... · Artificial Intelligence and the...

Artificial Intelligence and the future of the accounting profession: Déjà vu all over

again

Miklos A. VasarhelyiKPMG Distinguished Professor of AIS

November 15, 2019University of Toronto

1

AI AND THE SINGULARITY

The Singularity in artificial intelligence (AI) is the inflection point where machines advance beyond human intelligence and thus become self sufficient. Jon von Neumann first defined the term in the 1950s and it has been further advanced by well-known futurist and technologist Ray Kurzweil.

2

Ellen Glazerman – ED E&Y Foundation

1) We are preparing students for jobs that do not exist

2) Using technologies that have not been invented

3) To solve problems that have not been identified.Need AGILE

Processes

Outline

• The CarLab• What is intelligence• Evolution of Artificial Intelligence• Cognitive computing• Exogenous Data, • IPA• Estimates with machine learning

4

THE CARLAB

5

6

BRIGHAM YOUNG UNIVERSITYThe Ranking of Rutgers in the Accounting Areas

Areas Ranking 2008-2013 Ranking 2002-2013

Ranking 1990-2013

AIS #1 out of 179 #1 out of 207#1 out of 241

Audit #10 out of 320 #7 out of 370#11 out of 438

Financial #70 out of 356 #89 out of 406#83 out of 470

Managerial #120 out of 286 #80 out of 346#66 out of 413

Tax #53 out of 129 #76 out of 178#79 out of 246

Other #35 out of 171 #18 out of 248#25 out of 341

CarLab Analytic Research

Audit data analytics and

EDA

Process Mining at

Gamma Bank

Fraud Risk Assessment using EDA

Logit regression for

control risk assessment

Text Mining

Expert System for P-Card

Envisaging the future of audit and Big Data

Detecting duplicate records

Predictive Audit

Continuity equations

Exceptional Exceptions

Multidimensional clustering

for fraud detection

Continuity Equations at

HCA

Credit cardDefault

prediction

Rule-based selection for

transitory accounts

XBRL

ClientRetention

Project

MonitoringUnibanco’sbranches

XBRL InsuranceAnalytics

Litigationprediction

Visualization

InsuranceAnalytics

8

Predictive Analytics with Weather data

Choosing apps

CarLab

CognitiveDecision Aids

AI: Deep Learning

Robotic Process

Automation (RPA)

Intelligent Process

Automation (IPA)

Blockchain and Smartcontracts

Cluster Analysis of US

States

Recent research of PhD students

https://www.youtube.com/playlist?list=PLauepKFT6DK9vKn7-eKxzmxBegpe8v8xw

Name TitleAbdulrahman Alrefai Formalization of Internal Control Assessment: A Process Mining Application

Ahmad AlQassar Resisting Change in the Audit Profession: Two Case Studies from Multi-National Firms

Andrea Rozario Examination of Audit Planning Risk Assessments Using Verbal Protocol Analysis: An Exploratory Study

Cheng Yin Privacy-Preserving Information Sharing within an Audit FirmDeniz Appelbaum Using Drones in Internal and External Audits: An Exploratory Framework

Feiqi Huang Audit Evidence Index ProjectHe Li Are External Auditors Concerned about Cyber Incidents? Evidence from Audit Fees

Jiahua Zhou The Survived Companies with Going Concern Are Really Different from Those BankruptedJun Dai Towards Blockchain-based Accounting and AssuranceJun Dai Imagineering Audit 4.0

Zhaokai Yan Impact of Data Analytics on Managerial Accounting Using Balanced Scorecard FrameworkYunsen Wang An Application of Blockchain Technology to Fraud Detection

Yue Liu Risk Analysis Based on 10-K Item 1aTing Sun The Performance of Sentiment Features of 10-K MD&As for Financial Misstatement Prediction

Tiffany Chiu Apply Process Mining to Evaluate Internal Control Effectiveness AutomaticallyQiao Li Rule-Based Decision Support System for Audit Planning and Audit Risk Assessment

Lu Zhang Interactive Data Visualization for Error and Fraud Detection: Case Studies and Practice Implications

Content• Undergraduate, Graduate, PhD, & Audit Analytics Content

Undergraduate Graduate PhD Audit Analytics Certificate

• Introduction to Financial Accounting

• Introduction to Managerial Accounting

• Intermediate Accounting I• Intermediate Accounting II• Advanced Accounting • Auditing Principles• Management and Cost

Accounting• Accounting Information Systems• Business Law I• Business Law II• Federal Taxation I• Accounting in the Digital Era• Computer Augmented

Accounting• Decoding of Corporate Financial

Communications

• Accounting Principles and Practices

• Information Technology• Government and Not-for-

Profit Accounting• Advanced Auditing and

Information Systems• Advanced Accounting• Corporate Taxation• Income Taxation• Income Tax Estate and

Trust

• Special Topics in Accounting

• Survey of Accounting Information Systems

• Current Topics in Auditing• Machine Learning

• Introduction to Audit Analytics• Special Topics in Audit

Analytics• Information Risk

Management• Tutorials for Risk

Management

Digital Library

12

WHAT IS INTELLIGENCE?

What is intelligence?

• Is it being very cultured? – Being a renaissance person?

• Is it being able to solve problems no one solved before?

• Is it being Human?• Is it contingent on the moment in time?• Is it satisfying functionalities?

13

The Turing Test

Alan Turing: “Can Machines Think?”

Interrogator

ComputerHuman Being

Written Communication

Loebner prize established in 1990

15

EVOLUTION OF AI

• Dreyfus (1964) classifies traditional Artificial Intelligence (AI)work into four main areas: – game playing, – problem solving, – language translation, – and pattern recognition.

Evolution of AI

16

• In 1964, when Mr. McCarthy established the Stanford Artificial Intelligence Laboratory, the researchers informed their Pentagon backers that the construction of an artificially intelligent machine would take about a decade. Two decades later, in 1984, that original optimism hit a rough patch, leading to the collapse of a crop of A.I. start-up companies in Silicon Valley, a time known as “the A.I. winter.”

• Such reversals have led the veteran Silicon Valley technology forecaster Paul Saffo to proclaim: “never mistake a clear view for a short distance.”

17

When?

• In the early sixties Feigenbaum reoriented the work in AI by focusing not on basic paradigms and pure logical development but on the identification and formalization of human expertise often represented in the form of software systems. This led to the area of Expert Systems (ES) which became one of the five main areas of AI.

Evolution of AI

18

–Natural Languages, –Expert Systems, –Cognition and Learning, –Computer Vision and –Automatic Deduction.

Areas of AI 1990’S

19

• Expert Systems became the most popular area of AI and eventually the basis of many commercial, semi-commercial and prototype systems.

• Vasarhelyi, M. A. “Expert Systems in Accounting and Auditing,” in Artificial Intelligence in Accounting and Auditing, Vols. 1 to 6 , Markus Wiener Publishing Inc., New York. (1989 to 2002)

Evolution of AI

20

21

Expert Systems Literature• Shpilberg, D., Graham, L. E., & Schatz, H. (1986). ExperTAXsm: an expert

system for corporate tax planing. Expert Systems, 3(3), 136-151.• Shpilberg, D., and Graham, L. E. (1986). Developing Expertaxsm-An Expert

System for Corporate-Tax Accrual and Planning. Auditing: A Journal of Practice and Theory, 6(1), 75-94.

• Graham, L. E., Damens, J., and Van Ness (1991). Developing risk Advisor: An expert System for Risk Identification. Auditing: a Journal of Practice and Theory, 10(1), 69-96.

• Feigenbaum, E. A. (1981). Expert systems in the 1980s. State of the art report on machine intelligence. Maidenhead: Pergamon-Infotech.

• Buchanan, B. G., & Duda, R. O. (1983). Principles of rule-based expert systems. In Advances in computers (Vol. 22, pp. 163-216). Elsevier.

• Gray, G. L., Chiu, V., Liu, Q., & Li, P. (2014). The expert systems life cycle in AIS research: What does it mean for future AIS research?. International Journal of Accounting Information Systems, 15(4), 423-451.

22

• Although AI has been through cycles of hype and disappointment before, big technology companies have recently been scrambling to hire experts in the field, in the hope of building machines that can learn even more sophisticated tasks. (Economist, 2014)

23

Hype or disapointment?

IN ACCOUNTING AND AUDITING

24

Knowledge Base WorkingMemory

InferenceEngine

Inference Control

KnowledgeAcquisitionSubsystem

ExplanationSubsystem

User interfaceExpert or Knowledge Engineer

Expert Systems

25

System

OperationalReport

OperationalReport

OperationalReport

Filter

Database

System Operational Reports

Workstation

DF-level 0Alarm

Data Flow Diagrams

DF-level 1 DF-level 1 DF-level 1

DF-level 2

Reports Analytics Metrics

CPAS Architecture

26

–Natural Languages, –Expert Systems, –Cognition and Learning, –Computer Vision and –Automatic Deduction.

• And Now– Deep Learning / Cognitive Computing

Areas of AI 1990’S

27

Miklos A. VasarhelyiHelen Brown Liburd

Rutgers Business School

Exogenous data analytics for Auditing

Some sources

• Amazon sales• Google searches• Apps used• Calls made• GPS or JEEP location• Sites accessed• Car license plates photographed• Pictures of parking lots• Face recognition pictures• Site clickpaths

29

30

SocialMedia

Internet of

Things

Economicdata

Weather data

Locational data

Internet sales data

SearchData

Clickdata

ED may be of easier access

ED is likely less tamperable

ED relationships will be stochastic

ED is a form of confirmation

ED may complement many current procedures

ED may create many new procedures

Exogenous Data

Some other sources

• Security recordings of arrivals and departures of trucks from parking lots for assuring inventory changes

• Telephone records, associated with e-mails, to validate sales, ordering, and discrepancy determinations

• Examination of video streams in network TV to confirm that ads were actually placed. These can be linked to variations in order/ sales to validate the ad efficiency promised by ad agencies and marketing strategies

31

RADAR: external data that were mentioned in firm interviews

• Bloomberg data, Twitter (for estimating warranty liability)• Economic indicators• Capital IQ• Credit Ratings

32

The new data ecosystem: Cho, Vasarhelyi & Zhang

33

Developing A Cognitive Assistant For Audit Plan Brainstorming Sessions

Qiao Li

Rutgers Business School

QA

Architecture of the Proposed Audit Cognitive Assistant

Automatic Speech

Recognition

Query Classifier

Question or Action

text

Answer ShowAnswer

ExecuteAction

Action

Lucaindustry

Client

Position

Luca Luca

LucaRecommended Topics:General understanding, new events, business risks…

Processing…

You may also interested in:…

Query

Open an application

Inte

rface

Arch

itect

ure

Modules:

• Automatic Speech

Recognition (ASR)

• Language Understanding

• Dialogue Management

• Natural Language

Generation

• Text-to-Speech synthesis

Audit Related Applications It Can Access

Web Search

Open (ACL,

IDEA…)

Calculator

Open standards

Open templates

Audit workpaper

Calendar …….back

stag

es

uppo

rter

Knowledge Database

Knowledge about users

Knowledge Base

DBMS

Unstructured data

Domain Knowledge

35/35

Apply Deep Learning to Analyze Big Data for Predictive Auditing

Ting Sun & Miklos Vasarhelyi

Motivation

• AlphaGo beats European Go Champion.• Deep Learning1. Using a vast collection of Go moves from expert players(about 30 million moves in total)àtrained their system to play Go on its ownà as good as the best humans2. (self-reinforcing)matched the system against itself à generate a new collection of moves àtrain a new AI player that could be better than human

How could deep learning be used in audit?

Why deep learning?

• The future of Big data is deep learningthe biggest part of Big Data is the unstructured part, and it contains valuable information and learnable patternsimpossible and costly for human to extract features from unstructured data or label the structured data (whether there is a fraud or not)• Traditional way (i.e., SVM, LR) of applying machine learning to

predictive audit: need human to identify and detect features(attributes) of data and label the data

• Deep learning: The computer learns the inner structure and features of data itselfthe technology could mimic the human intuition and think like human brain and automatically developing new ways of representing the data （often based on ANN）

An illustration: face recognition

Apply deep learning to audit

Raw data:Mixture of

structured and unstructured data

Normal behavior

Irregularity against assertion A

Irregularity against assertion BIrregularity against assertion C

How to apply

• For a given assertion, auditor’s objective is to detect irregularities for this assertion:

• train huge volume of data (past data), including regular numerical data, semi-structured data, and unstructured data (video, audio, text)àMachine extracts features à find the patterns à generate the model(classifier)àauditors use the model to predict irregularities

• Reinforced learning: self-correction

Example: financial audit

• Step 1: train the data (historical records, video, audio, text)• Step 2: get the output : p(𝑓𝑟𝑎𝑢𝑑&I 𝑆𝑡𝑎𝑡𝑒*)• Step 3: calculate possible 𝑙𝑜𝑠𝑠&• Step 4: rank risk level based on p (𝑓𝑟𝑎𝑢𝑑&I 𝑆𝑡𝑎𝑡𝑒*) × 𝑙𝑜𝑠𝑠&• Step 5: take audit actions for top leveled risks (audit

recommendation system)• For the long term: reinforced learning à self-correction as the

state changes/new data are included

How to apply ?1. train huge volume of semi-structured or/and unstructured raw data

(e.g., video, audio, text data)2. Machine uses Deep Learning to extract features from the raw data

(within the black box) 3. find the patterns (characteristic types) 4. generate model (classifier) A5. auditors use model A to identify characteristic types from big data6. auditors combine information identified from last step with regular

structured data (like financial data) as audit evidence7. The past audit evidence can be used as training data to develop

model B (use supervised shallow learning) and use model B to predict frauds

8. As new data is collected, machine uses Reinforced Learning technique to improve the accuracy of model A (a continuous self-correction process)

Example: The Securities and Exchange Commission (SEC) comment letter• Step 1: collect and train the raw data, SEC comment letters

after reviewing registrants’ filings• Step 2: develop model A to extract features from data: words-

>phrases->sentences->paragraphs->meaning of entire letter (in this step, machine learns to understand the letter on its own)

• Step 3: use model A to classify different characteristic types of letters based on the meaning the machine understood in step 2

IPA

Intelligent Process Automation (IPA)

46

IPA

Artificial intelligence

Machine learning

Computer vision

Virtual agent

NLP/NLG Others

RPA

Other technologies

Big Data/ Analytics Drones IoT Others

IPA is “an emerging set of new technologies that combines fundamental process redesign with RPA and machine learning” (McKinsey, 2017).

“It is a suite of business-process improvements and next-generation tools that assists the knowledge worker by removing repetitive, replicable, and routine tasks.” (McKinsey, 2017).

Blockchain Smart workflow

Intelligent Process Automation (IPA)

47

(Adapted from Lacity & Willcocks, 2017)

IPA can cover the Automation Continuum

IPA

48

49

Intelligent Process Automation (IPA)

The “Sense-Think-Act” loop of IPA

50

(Adapted from UiPath, 2017)

The “Auditor-in-the-Loop” IPA Ecosystem

51

Accounting Estimates Using Machine Learning

Keshing Ding, SWUFEBaruch Lev, NYU

Xuan Peng, SWUFEMiklos A. Vasarhelyi, Rutgers University

November 21, 2019 Toronto

“Accounting estimates are pervasive in financial statements, often substantially affecting a company’s financial position and results of operations… ” (PCAOB 2018, p.3).

53

Accounting estimates

Accounting estimate examples:• fixed assets

• accounts receivable

• pension expenses and

incomes

Accounting estimates

• General Electric Example– 2016 net earnings is $8.2 billions.– Half came from a change in managers’ estimates.

“Contract assets increased $4,006 million in 2016, which was primarily driven by a change in estimated profitability within our long-term product service agreements …”

54

Improve estimates

• Causes of estimation errors– environment uncertainty– managers’ manipulation

• Machine learning– decreases manipulation: an independent, less-bias estimates generator– decreases uncertainty: take into account more factors in prediction

• Our Research– use machine learning algorithms to estimate losses for property & casualty

insurance companies– compare machine learning estimates with managers’ estimates

55

Research design

Business Line Year 0 Year 1 Year 2Private Passenger Auto Liability 40.64% 72.44% 86.76%Commercial Auto Liability 25.03% 50.74% 70.90%Workers’ Compensation 24.99% 56.11% 72.90%Commercial Multi-Peril 44.52% 69.22% 80.03%Homeonwer/Farmowner 72.62% 93.50% 96.83%

56

• Business lines (cumulative payment percentage)

• Training/Validation/Testing approach

Validate Train

Validate Train

Validate Train

Train Validate

Train Validate

Validate Train

Validate Train

Validate Train

Train Validate

Train Validate

Validate Train

Validate Train

Validate Train

Train Validate

Train Validate

Cross-validation results

57

Business line Sample Obs Accuracy Edge

Private Passenger Auto Liability

1996-2005 5949 12%1996-2006 6298 13%1996-2007 6602 26%

Commercial Auto Liability1996-2005 5383 42%1996-2006 5661 36%1996-2007 5957 37%

Workers’ Compensation1996-2005 4183 35%1996-2006 4398 43%1996-2006 4398 48%

Commercial Multi-Peril1996-2005 5235 33%1996-2006 5457 34%1996-2007 5846 42%

Homeowner/Farmowner1996-2005 6121 -12%1996-2006 6544 24%1996-2007 6946 24%

• The percent accuracy improvement of the ML loss estimates over managers’ estimates in 5-fold cross validation.

Holdout test results

58

Business line Sample Obs Accuracy Edge

Private Passenger Auto Liability

2006 670 26%2007 659 14%2008 637 37%

Commercial Auto Liability2006 620 20%2007 609 20%2008 592 49%

Workers’ Compensation2006 499 54%2007 498 55%2008 473 19%

Commercial Multi-Peril2006 582 50%2007 570 22%2008 563 -18%

Homeowner/Farmowner2006 697 51%2007 692 38%2008 678 52%

• The percent accuracy improvement of the ML loss estimates over managers’ estimates in holdout test.

Conclusion

• Accuracy edge: accounting estimates generated by machine learning are potentially superior to managerial estimates.

• Benchmark: estimates generated by machine learning can be used by managers and auditors as benchmarks against which managers’ estimates will be compared. Large deviations will suggest a reexamination of managers’ estimates.

• Potential: machine learning could be used to generate estimates to be report in the first place.– enhance the reliability (no manipulation) and consistency of accounting

estimates.

59

The FASB could

• Create a machine learning estimate for a very narrow industry corresponding to reporting lines of business– Determine estimate based on an allocated percentage or and

adjusted percentage of the business• Allow businesses to do their computations and estimates

with– A pre-set estimation methodology with machine learning or the

machine learning done by the standard setter– The inputs to the estimation methodology (variables) be

auditable values

60

CONCLUSIONS

61

Computerization of Occupations

62

Adapted from: “The Future of Employment: How Susceptible are Jobs to Computerisation?” (Frey and Osborne, 2013)

Conclusions

• Disruption will come first from exogenous data and then from cognitive assistants

• AI will take longer to really affect• External validation is a new way of thinking• Labor replacement will come mainly from RPA• Technological process retrofitting will be necessary on the

audit process and in standards• The purchase/ acquisition of data will require substantive

resources but will not depend on the client IT personnel• Auditors will also create collection needs for IoT

64